Technology & privacy
Parakeet on a Mac, explained without the jargon
A family of speech models from NVIDIA, designed for speed at good accuracy. That design goal is why several Mac dictation apps use it for English rather than a larger model.
Last updated
What it is
Like Whisper, Parakeet is a set of trained model files you can download and run. It is published by NVIDIA and comes in several variants, generally compact, with strong English performance.
The distinguishing property for our purposes is throughput. Parakeet models were built with inference speed as a first-class goal rather than as a consequence of being small, and on Apple Silicon that shows up as dictations returning quickly enough that you do not register a wait.
Why speed is the right thing to optimise for dictation
This is the part worth understanding, because it is counterintuitive if you think of speech recognition as a transcription problem.
Dictation is an input method. The comparison in the user's mind is not "how accurate was this?" but "was this faster than typing?". A model that is half a point more accurate and takes an extra two seconds loses that comparison every time, because two seconds of staring at nothing is a longer interruption than one wrong word.
| Live dictation | File transcription | |
|---|---|---|
| Latency tolerance | Under a second | Minutes are fine |
| What a delay feels like | The tool is broken | The tool is working |
| Best model choice | Fast and compact | Large and accurate |
| How errors are fixed | By you, immediately | In an editing pass later |
| Typical Parakeet fit | Very good | Usable |
| Typical large-Whisper fit | Sluggish | Very good |
Where Parakeet is weaker
- Language coverage is narrower than Whisper's larger variants. If you dictate in a less widely supported language, this is decisive.
- Difficult audio still favours a larger model. Noisy rooms, strong accents in a second language, overlapping speakers.
- It is not a transcription workhorse. For an hour of recorded audio where latency does not matter, a larger model is the better tool.
- The hardware matters. Its speed advantage is most pronounced on Apple Silicon; on older hardware the picture is different.
How it is used in practice
In VV, Parakeet and Whisper-family models both ship and the app selects based on your Mac and language. The model is downloaded on first run, verified by checksum before it is ever loaded, and stored in the app's own data directory.
There is no general model picker in the standard interface. That is a deliberate narrowing — and a legitimate reason to prefer an app like Superwhisper if choosing the model yourself is something you want.
Does the model name matter when choosing an app?
The one case where it does matter: your language. If a model family does not cover the language you dictate in, no amount of good engineering around it will help. Check that first and treat everything else as secondary.
Questions
Is Parakeet better than Whisper?
Not better — different. Parakeet is designed for speed at good accuracy, which suits live dictation. Whisper's larger variants are broader across languages and more robust on difficult audio, which suits transcription. Neither dominates.
Does Parakeet run on Apple Silicon?
Yes, and that is where its speed advantage is clearest. Unified memory removes the copy that slows local inference on other architectures.
Is it free?
The models are openly available, so an app can run them on your machine at no per-use cost. That is part of why local dictation apps can be sold as a one-off licence rather than a subscription.
Can I choose Parakeet in VV?
VV ships Parakeet and Whisper-family models and selects for your Mac and language rather than exposing a general picker. If choosing yourself matters, Superwhisper is more flexible and that is a fair reason to prefer it.
Does it support languages other than English?
Coverage is narrower than Whisper's larger variants. If you dictate in a less widely supported language, check specifically — Apple's built-in dictation or a cloud service may serve you better.